Title : Artificial intelligence in the auxiliary diagnosis of hepatocellular carcinoma
Abstract:
Objective: To systematically review the application progress of artificial intelligence (AI) in imaging diagnosis, pathological diagnosis and tumor grading of hepatocellular carcinoma (HCC), grade and comment on the existing evidence according to the research design characteristics, analyze the clinical value and existing limitations of AI in the auxiliary diagnosis of HCC, prospect the future development direction, and provide a reference for the precise diagnosis and treatment of HCC.
Materials and Methods: With hepatocellular carcinoma, artificial intelligence, deep learning and radiomics as the core keywords, the relevant domestic and foreign research literatures in the past 5 years were retrieved. The application studies of AI in ultrasonic/CT/MRI imaging diagnosis, pathological image recognition and biomarker prediction, and non-invasive tumor grading of HCC were sorted out. According to the three core indicators of sample size, multicenter or not, and independent external verification or not, the included studies were graded as high/medium/low evidence level, and the application advantages, methodological characteristics and clinical transformation obstacles of AI technology in the auxiliary diagnosis of HCC were systematically analyzed.
Results: AI technology could significantly improve the detection rate and differential diagnosis accuracy of HCC lesions, and showed unique value in clinical diagnosis difficulties such as alpha-fetoprotein (AFP)-negative HCC and small HCC (≤2cm). High-evidence studies (multicenter + external verification) showed that the overall accuracy of AI models in HCC imaging diagnosis reached 83%~96%, the area under the curve (AUC) of pathological image diagnosis was ≥0.949, and the AUC of multimodal fusion models in small HCC subgroup reached 0.996. AI models based on ultrasound/MRI could realize non-invasive preoperative grading of HCC, and the AUC of ultrasound radiomics model combined with clinical features for grading prediction reached 0.849. The existing studies still had obvious limitations: most of them were single-center and retrospective studies with only internal verification, resulting in insufficient model generalization ability; deep learning models had the characteristic of "black box" with poor interpretability; the non-uniformity of image acquisition parameters and reference standards led to low model reproducibility; at the same time, they faced transformation obstacles such as data privacy, algorithm bias and poor clinical adaptability.
Conclusion: AI has shown excellent clinical application potential in the field of HCC auxiliary diagnosis. Technological innovations such as multimodal fusion, interpretable AI and super-resolution reconstruction have promoted its evolution from a simple image auxiliary tool to a clinical decision support system. In the future, it is necessary to focus on carrying out large-scale, prospective and multicenter studies, improve the external verification system of models, promote the interpretability of algorithms and the standardization of methodology, and push the deep integration of AI with the clinical diagnosis and treatment process of HCC, so that it can play a greater clinical value in the early screening, precise diagnosis and individualized treatment of HCC.
Key words: Hepatocellular carcinoma; Artificial intelligence; Deep learning; Radiomics; Auxiliary diagnosis
